This is a basic feature of the str product, easiest to see in bland QM in phase space. (Just take D=2 dimensions x and p, and $\theta_{ij} =\hbar \epsilon_{ij}$.) You might, or might not profit from the Real Mc Coy: Zurek, W. H. (2001). Sub-Planck structure in phase space and its relevance for quantum decoherence, Nature 412(6848), 712-717.
The integral representation of the *-product (2.17) you are discussing (Baker 1958; Srinivasan & Wolf 1975) are eqns (15) & (163) of our revised tutorial booklet. Basically, in words, the *-product entangles translations of f and g and the smaller the ambit fo the functions, the larger the gradients involved, so the bigger the translations (non locality). This is all mere verbiage, of course, unless it summarizes basic paradigm calculations you could perform yourself, like your author's (3.8)--a constant, infinitely broad! (Exercise 0.4 in our booklet).
An easy sketch for this (prove it!) is our (rapidity composition) Corollary (82), namely $$ \exp( -a(x^2+p^2)/\hbar ) \star \exp( -b(x^2+p^2)/\hbar )=\exp\left ( -{(a+b)(x^2+p^2)\over \hbar( 1+ab)}\right ). $$ Take both (squared) widths of the 2d Gaussians composed to be small: $\hbar/a=\hbar/b=\hbar\delta$, so their star product has a width $(\delta + 1/\delta)\hbar/2$, effectively a huge width ℏ/2δ. As δ goes to zero to get to the Dirac δ-function limit, the width of the respective star product goes to infinity!
To flesh out the initial verbiage paragraph then, the "Bopp shift" representation of the star product is $$f(x,p)\star g(x,p)=f\left(x+\frac{i}{2}\partial_p,~p-\frac{i}{2}\partial_x\right) g(x,p), $$ where the gradients only act on g, not f. But if g is confined to a small ambit, the gradients are effectively huge, so the shifts of the arguments of f are huge, and the non-locality augments dramatically. You can dress this up with symplectic Fourier transform verbiage, but it is a simple, primitive phenomenon.